Automated Nighttime Fog Detection and Masking Using Machine Learning from Near Real-Time Satellite Observations

Fog significantly reduces visibility, impacting transportation and safety, particularly in regions like the United Arab Emirates (UAE) where it is a regular
occurrence, in particular in the winter months. This study develops a machine learning-based approach for automated fog detection and masking from near real-time observations from the Spinning Enhanced Visible and Infrared Imager (SEVIRI) instrument onboard the Meteosat Second Generation spacecraft to enhance fog detection and forecast. We evaluated six basic machine learning (ML) models trained with four different methods: (1) supervised training using SEVIRI pixel data and fog observations over airport stations; (2) as approach (1) but incorporating infrared channel data; (3) training with labeled fog and no-fog regions identified in SEVIRI night microphysics Red-Green-Blue (RGB) images through k-means clustering; and, (4) a fusion approach combining station-labeled data (approach 1) and k-means clustered-labeled data (approach 3). Among the models, the eXtreme Gradient Boosting (XGBoost) demonstrated slightly higher performance. Models trained on station data (approach 1) achieved a Probability of Detection (POD) of 0.73 and a False Alarm Ratio (FAR) of 0.11. For spatial fog masking, models trained on a combination of station-labeled and k-means cluster-labeled data (approach 4) performed best. Overall, the XGBoost method and the fusion approach (4) are recommended for fog detection and masking in the hyper-arid UAE. These findings demonstrate the potential for trained ML models to deliver accurate, near real-time fog detection and masking, enhancing monitoring over broad areas.